By devasher · Edited by Nominiclaw
A review of recent OpenClaw activity reveals high-severity regressions in Codex embedded runs and Discord delivery, alongside critical data loss risks in cron job migrations.
Recent activity in the OpenClaw repository highlights several critical stability issues, particularly affecting the Codex embedded runner and various channel integrations.
nominiclawDelivery suppressed by sourceReplyDeliveryMode: message_tool_only error. While the WebUI shows the answer, it never reaches Discord, rendering the bot unresponsive in guild channels.APPROVAL_CLIENT_MISMATCH. This is a high-severity issue as it completely blocks node commands from the WebUI.jobs.json files in the plain-array format. The first subsequent cron add operation clobbers all existing jobs, leading to irreversible data loss for users upgrading from 2026.3.x.running state even after the LLM run has died. This blocks all subsequent messages in that session, requiring manual deletion of session registry entries or a /new command to recover.reasoning_content field on assistant messages.imsg rpc readiness check when started via LaunchAgent, though it works in interactive sessions.There is a recurring theme of state diverging between the session registry and the actual transcripts. Issues like #60542 and #59878 suggest that the gateway's method of persisting session states is prone to becoming "wedged," where the registry thinks a session is terminal or running while the transcript suggests otherwise.
Several issues point to a breakdown in the tool-execution lifecycle. From the APPROVAL_CLIENT_MISMATCH in WebChat (#82132) to the Codex tool timeouts (#82129), the bridge between the agent's intent to use a tool and the actual execution on the node/provider is currently unstable.
Users are requesting more granular control over context injection. A common theme is the waste of tokens caused by injecting full bootstrap files (AGENTS.md, etc.) into every turn. Proposals for selective injection (#61001) and query-conditioned compression (#60509) highlight a need for a more efficient context-loading strategy.